Imagine, in five to 10 years’ time, billions of IoT devices will be live in our world, connected to many different mobile networks through several radio technologies. Not only will we have 5G, 4G, 3G, NB-IoT, LPWA and WiFi as radio access technology standards for IoT, every one of them will have many possible different setups and configurations in commercial networks. Thus, the way – or what we call it ‘language’ – that IoT devices will communicate with their serving network will be fragmented and will have many ‘dialects.’ The ‘languages’ that we are talking about here is essential for IoT applications, as these are the signaling messages in a fine-designed protocol that make the connectivity reliable, robust, secure and smart.

Given the fact that the ‘language’ talked in the future IoT world is complex and fragmented, we can’t help to ask ourselves one question. What kind of IoT cloud system can understand the dynamic behavior of billions of IoT devices and orchestrate the world of these things? The answer is that the system needs to be powered by Artificial Intelligence (AI), and the dynamic device behavior needs to be learned by Machine Intelligence.

Thanks to advanced modern GPU´s (graphics processing units) and the amount of data that we can collect nowadays, deep learning on large-scale neural networks becomes possible and has already brought business value in many applications such as face recognition, natural language programming, e-commerce recommendation etcetera. Why does commercialized AI first happen in these fields? The answer is that a large amount of data is essential for deep learning model’s performance. The information behind these applications come from the digital world and in a shape of digital data that can easily be used as training data for neural networks. However, data such as pictures, human language, text, weblogs and such are only small portions of the total information from human activity. With the help of IoT, more data from the physical world can be collected, and next wave of commercialization of AI technology will for sure happen on machine learning with IoT data. LSTM (Long Short-Term Memory) is one kind of recurrent neural network that draws a lot of attention in the data science community today. It shows promising performance by memorizing the dependency and pattern between data points in sequential data. We have seen significant improvement on machine translation, transcription and OCR (Optical Character Recognition) on human language in the last two years, all thanks to training innovation on large recurrent neural network models, with cells similar to LSTM.

In our deep learning practice, we see more than 85% prediction accuracy on IoT device behavior after training an LSTM model on device ‘language’ data. The experiment on live IoT device behavior data shows that machine learning can help us understand device ‘language’ in a smart and scalable way. Just like natural human language programming – once you have the trained model representing the pattern and logic of this language, many applications can be built on top. Device behavior prediction, anomaly detection, and smart connectivity are the cool applications that we expect from modeling device ‘language’ using machine learning.

At Tele2 IoT, we truly believe AI will change the way of managing a massive number of devices and, that deep learning backed engine in our cloud system will help us understand the ‘language’ talked between device and network in depth. It will help us orchestrate the smarter world – the IoT world.